International Conference on Advances in ICT for Emerging Regions (ICTer 2020): 300 – 301

Abstractive Web News Summarization Using

Knowledge Graphs

Abstract— The rapid progresses in digital data acquisition

techniques have led huge volume of news data available in the

news websites. Most of such digital news collections lack

summaries. Due to that, online newspaper readers are

overloaded with lengthy text documents. Also, it is a tedious task

for human beings to generate an abstract for a news event

manually since it requires a rigorous analysis of the news

documents. An achievable solution to this problem is condensing

the digital news collections and take out only the essence in the

form of an automatically generated summary which allows

readers to make effective decision in less time. The graph based

algorithms for text summarization have been proven to be very

successful over other methods for producing multi document

summaries. The summary generated from knowledge graphs is

more in line with human reading habits and possesses the logic

of human reasoning. Due to the fast growing need of retrieving

information in abstract form, we are proposing a novel

approach for abstractive news summarization using the

knowledge graphs to fulfill the need of having more accurate

automatic abstractive text summarization.

Keywords—News Documents, Abstractive Summarization,

Knowledge Graphs, Data Mining, Natural Language Processing

I. INTRODUCTION

Within the recent years, we have seen extraordinary

interest in news aggregation and browsing. Hence news

websites becoming increasingly popular. Usually the Internet

provides excessive quantity of information for every news

topic than what is needed. Most of such electronic news

collections lack summaries. Hence online newspaper readers

are overloaded with lengthy text documents where smaller

version would do. Extracting knowledge manually from a

collection of newspapers may cause confusions, miss the track

for user and time consuming. Due to this, the selection of the

best collection of information for a particular news topic in a

minimum conceivable time is still a challenge. Most of the

existing applications for text summarization are extractive and

only few applications are generate abstractive summaries. The

graph based algorithms for abstractive text summarization

have been proven to be very successful over other methods for

producing multi document summaries. The summary

generated from knowledge graphs is more in line with human

reading habits and possesses the logic of human reasoning.

Therefore, we propose a feasible solution to this problem by

assembling the key information of a news event published in

news websites and taking out only the essence in the form of

an automatically generated summary which allows

newsreaders to make effective decision in less time.

II. LITERATURE REVIEW

Automatic text summarization techniques such as cluster

based, template based, ontology based, semantic graph based

methods, machine learning, fuzzy logic, neural network

approaches proven to be very useful over 50 years up to now

[1],[2]. Automatic text summarization approaches are two

fold as extractive and abstractive summarization [1]–[4].

Extractive summarization techniques extract important

sentences or phrases from the original text documents based

on the weight of statistical and linguistic features and produce

a summary without changing the original text [1]. Abstractive

summarization generates the summary by understanding the

source text using linguistic methods and generate new

sentences by improving the focus of a summary while

reducing the redundancy rate [1], [2], [5]. Abstractive

summarization is more complex because it requires deeper

analysis of source document(s) and concept-to-text generation

[4]. These techniques can be further classified as structured

and semantic based approaches [2], [5]. Semantic based

approaches provide more coherent, information rich and well-

structured abstractive summary than structured based

approaches. Due to the semantic representation, semantic

graph based approaches have been proven to be very

successful over other methods for producing multi document

abstractive summaries[6]. According to the Gupta et al. [7] the

semantic-based methods are not able to achieve similar

performance to deep learning approaches. Machine learning

approaches to generate automatic summaries [7]-[10] have

greatly improved compared to the primitive text

summarization methods, but they cannot be combined with

the background information to obtain higher level abstraction

[10] and cannot scale to the requirement of large data sets. In

order to understand a text well enough, we need some

background knowledge. Wu et al. [10] and Kalpa et al. [11]

inspired by those ideas and used knowledge graph to generate

automatic text summaries. Knowledge graphs capture domain

specifics and add rich and explicit semantics to infer

additional knowledge and encapsulate a large amount of

knowledge for human and machine consumption [11].

Graph based approaches use ranking functions comprising

of one or more sentences weighting features [12],[13] to

determine the relevancy of each edge for inclusion into the

summary. According to the literature, we have identified that

traditional word co-occurrence measures like TF-IDF gives

more importance to the words that are more frequent in the

document. Hence, a pure ranking algorithm can be less

effective because ranking alone cannot determine what kinds

of features are selected for the summary. But, increasing the

set of attributes to include semantic properties and topological

graph properties in the ranking function yields statistically

significant improvement for the abstractive summaries [12],

[13]. Knowledge graphs lead to information overload, and

hence proper summarization techniques need to be explored

[11]. Techniques in data mining are widely used to explore

data from various perspectives and uncovering unexpected

relationships between pieces of text and summarizing it into

useful information. Therefore, computers use text mining to

discover valuable and interesting information and knowledge

techniques. Among these techniques, Association Rule from

unstructured textual data in large volumes using different

Mining (ARM) can be used in discovering the relationships

M.V.P.T.Lakshika

University of Colombo School of

Computing

Colombo, Sri Lanka

tlv@ucsc.cmb.ac.lk

H.A.Caldera

University of Colombo School of

Computing

Colombo, Sri Lanka

hac@ucsc.cmb.ac.lk

W.V.Welgama

University of Colombo School of

Computing

Colombo, Sri Lanka

wvw@ucsc.cmb.ac.lk

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:30:24 UTC from IEEE Xplore. Restrictions apply.

M.V.P.T.Lakshika, H.A.Caldera, W.V.Welgama 301

International Conference on Advances in ICT for Emerging Regions (ICTer 2020) 05

– 06

November 2020

for decision making process and discovering correlations

between multidimensional features in the text documents [14],

[15].

III. RESEARCH PURPOSE

We are proposing a novel approach to bridge the knowledge

gap between Natural Language Processing and Data Mining

fields to generate more cohesive, readable abstractive web

news summarization using the knowledge graphs. This

approach overcome the downsides in existing abstractive

summary generation and enrich the accuracy of sentence

ranking function using the knowledge derived from

association rules in data mining for generating a better

abstractive summary. According to the literature, any existing

applications do not generate abstractive summaries on

multiple dimensions (topics) in news articles and also do not

generate ‘Updated news summaries’ for an existing

abstractive summaries. The proposed approach generate both

abstractive summaries on multiple dimensions or topics and

update summaries to help readers to read and track news

updates very easily.

IV. PROPOSED METHODOLOGY

As shown in the figure 1, after the text pre-processing steps,

ARM is applied to the automated indexes generated from

phase 1, and derived all the significant association rules from

the phase 2. In the phase 3, knowledge graph is formed with

two layers. Domain specifics in the newspapers are captured

in the data layer and semantic layer adds rich and explicit

semantics on top of the data layer to infer additional

knowledge using ontologies and lexical database such as

WordNet. A novel ranking algorithm which includes

linguistic and semantic features in the news documents along

with the knowledge mined from association rules in the phase

2 will be used to select the top-k entities from the knowledge

graph. In the phase 4, abstractive summaries are generated

using the SimpleNLG language generator. Updated

abstractive summaries are generated for an existing

abstractive summary. Generated correlations between

multidimensional features in the news documents are used to

generate multidimensional abstractive summaries. The

collection of news documents released by Document

Understanding Conference (DUC) will be used as the

experimental dataset. In the phase 5, human evaluation using

the domain experts will be carried along with the intrinsic

evaluation techniques such as recall, precision, F1 measure,

Pyramid and ROUGE evaluation.

REFERENCE

[1] R. Chettri, “Automatic Text Summarization,” Int. J. Comput. Appl.,

vol. 161, no. 1, p. 3, 2017.

[2] N. Bhatia and A. Jaiswal, “Trends in Extractive and Abstractive

Techniques in Text Summarization,” 2015, doi: 10.5120/20559-2947.

[3] A. Patil, “Automatic Text Summarization,” Int. J. Comput. Appl., vol.

109, no. 17, p. 2, 2015.

[4] N. R. Kasture, N. Yargal, N. Singh, N. N. Kulkarni, V. Mathur, “A

Survey on Methods of Abstractive Text Summarization,” 2014.

[5] D. K. Gaikwad and C. N. Mahender, “A Review Paper on Text

Summarization,” vol. 5, no. 3, p. 7, 2016.

[6] C. S. Yadav, A. Sharan, and M. L. Joshi, “Semantic graph based

approach for text mining,” in 2014 International Conference on Issues

and Challenges in Intelligent Computing Techniques (ICICT),

Ghaziabad, India, 2014, pp. 596–601, doi:

10.1109/ICICICT.2014.6781348.

[7] S. Gupta and S. K. Gupta, “Abstractive summarization: An overview

of the state of the art,” Expert Syst. Appl., vol. 121, pp. 49–65, May

2019, doi: 10.1016/j.eswa.2018.12.011.

[8] P. Kouris, G. Alexandridis, and A. Stafylopatis, “Abstractive Text

Summarization Based on Deep Learning and Semantic Content

Generalization,” in Proceedings of the 57th Annual Meeting of the

Association for Computational Linguistics, Florence, Italy, 2019, pp.

5082–5092, doi: 10.18653/v1/P19-1501.

[9] Z. H. Ali, “Multi-Document Text Summarization using Fuzzy Logic

and Association Rule Mining”, University of Technology –

Department of Computer Science, no. 41, p. 19, 2018.

[10] P. Wu, Q. Zhou, Z. Lei, W. Qiu, and X. Li, “Template Oriented Text

Summarization via Knowledge Graph,” in 2018 International

Conference on Audio, Language and Image Processing (ICALIP),

Shanghai, Jul. 2018, pp. 79–83, doi: 10.1109/ICALIP.2018.8455241.

[11] K. Gunaratna, “Semantics-based Summarization of Entities in

Knowledge Graphs,” p. 146, 2017.

[12] M. Subramaniam and V. Dalal, “Test Model for Rich Semantic Graph

Representation for Hindi Text using Abstractive Method.,” vol. 02, no.

02, p. 4, 2015.

[13] I. F. Moawad and M. Aref, “Semantic graph reduction approach for

abstractive Text Summarization,” in 2012 Seventh International

Conference on Computer Engineering & Systems (ICCES), Cairo,

Egypt, 2012, pp. 132–138, doi: 10.1109/ICCES.2012.6408498.

[14] C. E. Crangle, “Text Summarization in Data Mining,” in Soft-Ware

2002: Computing in an Imperfect World, vol. 2311, D. Bustard, W.

Liu, and R. Sterritt, Eds. Berlin, Heidelberg: Springer Berlin

Heidelberg, 2002, pp. 332–347.

[15] H. Mahgoub, D. Rösner, N. Ismail, and F. Torkey, “A Text Mining

Technique Using Association Rules Extraction,” vol. 2, no. 6, p. 8,

2008.

Figure 1: Proposed Methodology

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:30:24 UTC from IEEE Xplore. Restrictions apply.